This monthly report analyzes the Peak Shaving system of Fort William TS power station during 2021-March. This report contains four sections: First, the graphical results of peak shaving activities of power consumption. Second, the numerical results peak shaving activities of power consumption and expected bill saving. Third, the monthly performance of energy forecasting models. Four, detailed peak shaving activities of the highest five peak days during this month.
Figure 2.1 shows a monthly forecasting graph. The black line describes the actual power consumption curve; the red line describes the prediction curve and the yellow line describes the battery discharging period set by the predicted peak.
Figure 2.1: Monthly forecasting graph
Figure (2.2) shows a comparison of the expected power consumption curve(black) and after the peak-shaving curve(red). According to figure (2.2), The highest peak of power consumption of the month remains 59443.2kW (2021-03-01 19:55:00). Figure (2.3) shows the detailed peak shaving result of the highest peak of the month which is in 2021-03-01.
Figure 2.2: Monthly Peak Shaving Activies graph
Figure 2.3: Highest peak day
As shown in Table(3.1), The peak of power consumption remains 59443.2kW occur at 2021-03-01 19:55:00. After the peak-shaving activities, the highest peak is not been reduced. The monthly energy purchasing cost is $29721.6. The detailed situation is in Section 2: The results of Peak Shaving.
| Parameter | March Expected Peak | March Shaved Peak | Total Reduction |
|---|---|---|---|
| Power Consumption(kW) | 59443.2 | 59443.2 | 0 |
| Billing($) | 29721.6 | 29721.6 | 0 |
In this energy forecasting system, we combine four different models. Table (4.1) show the monthly performance of each model. There are four machine learning algorithms in total: Cubist, Xgboost, (feedforward) Neural Network and LSTM (Long short-term memory). The ensemble model combines all four algorithms and battery discharging events base on the ensemble model.
Relative root mean square error (rRMSE), which is RMSE divided by the average power consumption of tested day, between 1:00 pm to 12:00 pm to represent the performance of a model on the ability of prediction peak period. In the following formula, n represent number of data between 1:00 pm to 12:00 pm, yi and xi are the prediction and real power consumption receptively.
\[\begin{align} rRMSE = \frac{1\sqrt{(\frac{1}{n})\sum_{i=1}^{n}(y_{i} - x_{i})^{2}}}{\frac{1}{n}\sum_{i=1}^{n}x_{i}} \end{align}\]
Average Peak time error represent the mean of daily time different of expected peak time and predicted peak time. Pecentage of daily peak shaved represent the pecentage of days which discharging period coverd actual peak time in this month. For example, 0.8 mean this peak shaving system successfully shaved 80% of daily peak in this month.
| Parameter | Cubist | Xgboost | Random Forest | Nerual Network | LSTM | Ensemble System |
|---|---|---|---|---|---|---|
| rRMSE | 0.10 | 0.09 | 0.09 | 0.10 | 0.10 | 0.08 |
| Average Peak time error | 44.68 | 39.35 | 29.52 | 26.13 | 20.16 | 27.48 |
| percentage of daily peak shaved | 0.71 | 0.68 | 0.84 | 0.87 | 0.90 | 0.94 |
The accuracy of energy forecasting on the highest peak of the month is an important factor to measure the performance of each model. Figure (4.1) show the performance of each model on the day with highest peak.
Figure 4.1: Daily Models Performance